A SURVIVAL ANALYSIS AND GRADE PREDICTION MODEL FOR LUNG SQUAMOUS CELL CARCINOMA BASED ON MULTIPLE-INSTANCE LEARNING AND MULTI-SCALE TRANSFORMER
LE LI, Yong Liang, YONGQIANG GUO, Mingwen Shao, SHANGHUI LU, Shuilin Liao · Journal of Mechanics in Medicine and Biology · 2025
Respiratory diseases are now the third leading cause of death worldwide. Lung squamous cell carcinoma (LUSC) has one of the highest morbidity and mortality rates among respiratory diseases. Therefore, constructing a model based on the pathological images for survival analysis and grade prediction of LUSC is of great significance for designing personalized solutions for LUSC. The current LUSC survival analysis and grade prediction model mainly has two issues. First, it lacks high-performance multi-scale feature methods, limiting the model’s ability to discriminate LUSC case images. Second, the feature extractor has a high degree of redundancy, resulting in many repeated calculations. Excessive feature extraction can be regarded as noise, which not only increases the training cost of the model but also reduces the model performance. This study aimed to address these limitations by proposing a multiple-instance learning-based multi-scale transformer (MSTrans-MIL) for LUSC survival analysis and grade prediction. The contributions of this study were as follows. First, we proposed a feature sampling module (FSM) based on a self-attention mechanism, which was conducive to reducing information redundancy in the input space and improving the model’s applicability. Second, we constructed a multi-scale pathology feature extraction module based on self-supervised learning and introduced a convolution–transformer to adequately extract the local and global features of images on different scales. The multi-scale chains are also beneficial to understanding the interaction between the tissue microenvironment and tumor cells. In addition, a multi-scale feature encoder with sparse Transformer was proposed to further reduce the feature redundancy, and a multi-scale feature aggregation module using the gating unit was constructed to enhance the hierarchy of the feature representations and improve the robustness and accuracy of the model. Abundant ablation and comparison experiments demonstrated that the proposed MSTrans-MIL could reduce feature redundancy and improve the prediction of LUSC grading and prognosis.